Neural Network Handoff Algorithm in a Joint Terrestrial-HAPS Cellular System

164 ECTI TRANSACTIONS ON ELECTRICAL ENG., ELECTRONICS, AND COMMUNICATIONS VOL.3, NO.2 AUGUST 2005

Neural Network Handoff Algorithm in a JointTerrestrial-HAPS Cellular System

Raungrong Suleesathira and Sunisa Kunarak, Non-members

ABSTRACT

Handoff algorithm is used in wireless cellular sys-tems to decide when and to which base station (BS)to handoff in order that the services can be continueduninterrupted. In this paper, we propose a handoffalgorithm based on the neural network in a joint sys-tem of terrestrial and high altitude platform station(HAPS) cellular systems. As a revolutionary wirelesssystem, HAPS can supply services for uncovered areaimproving total capacity of service-limited area by aterrestrial BS. Radial-Basis function (RBF) networkis used for making a handoff decision to the chosenneighbor BS. The neural inputs consist of the aver-aged signal strength received from the serving andnearby BSs, user directions estimated by the MUSICalgorithm on an antenna array, and traffic intensi-ties. Positioning a mobile station (MS) is obtainedby apply the timing advance (TA) concept. Perfor-mance comparisons among the presented method andusing the backpropagation (BP) neural network andthe conventional Hysteresis method are given in formsof (1) handoff rate, blocking rate, dropping rate at theacceptable grade of service and (2) the difference be-tween the signal power radiated by the serving BSand the minimum required received signal power be-fore call dropping for both Hata model and Rayleighfading.

Keywords: Handoff, Radial-basis function net-works, Backpropagation networks, MUSIC, Timingadvance, Rayleigh fading.

1. INTRODUCTION

In mobile communications, the continuity of com-munication without terminating an ongoing call orblocking new calls is very crucial to enhance highquality of cellular services. Handoff algorithm makesit possible to maintain link quality. In Hysteresismethod [1], handoff occurs when the difference of sig-nal strength received from target and current BSs ishigher than Hysteresis level. Because of fading ef-fect, the difference can be fluctuated for brief periodsof time which results in unnecessary handoffs. Such

CM1R38: Manuscript received on August 28, 2004 ; revisedon June 25, 2005.

The authors are with the Department of Electronic andTelecommunication Engineering King Mongkut’s University ofTechnology Thonburi, Tungkru, Bangkok, 10140 Thailand.E-mail: [email protected]

back and forth handoff is known as the ping-pongeffect. In addition to network resource waste, callsmight be terminated if decision delay is long due tothe high Hysteresis level.

Recently, neural networks have been utilized to im-prove handoff algorithms due to its ability to handlelarge data in fast processing. Adaptive parameterssuch as user speeds, received signal strengths for pat-tern classification provide a multiple of criteria hand-off algorithm [2]. Neural network is trained to pre-dict a user’s transfer probabilities which represent theuser movements [3]. A technique to recognize signalpatterns of a MS using probabilistic neural networkis introduced in Rayleigh fading channel [4]. Usingstatistical pattern recognition of signal strength [5-6]can improve the efficiency of handoff algorithms.

In this paper, we present an effective handoff algo-rithm based on radial basis function (RBF) networks[7] in a joint terrestrial-HAPS cellular system [8]. Fig-ure 1 shows the conceptual model of the joint cellu-lar system. The inputs of neural network depend onsignal strengths, mobile directions and traffic inten-sities, which are used to make a handoff decision tothe chosen adjacent cell or HAPS. We use an antennaarray with the MUSIC [9] (MUltiple SIgnal Classifi-cation) to estimate directions of arrival of MS signals.The signal strengths of mobiles are computed in theHata model and Rayleigh fading [10]. HAPS cellu-lar system can be considered as a complementary tothe terrestrial cellular system, to improve and expandthe coverage services. As shown in Fig. 1, HAPS cansupply services to the mobile having weak signal fromthe serving terrestrial BS influenced by shadowing,turning corner as well as being outside the terrestrialcoverage. The timing advance (TA) concept [11-12] inGSM system and the power of mobile signal are usedfor checking if the mobile are in an obstacle area.

This paper is organized as follow. Section 2 re-views the Hysteresis method. Handoff decision basedon RBF network is presented in section 3 and basedon the backpropagation network is next presented insection 4. The MUSIC algorithm is explained in sec-tion 5. The concept of timing advance is in section6. The handoff algorithm and HAPS improving thecall maintenance are presented in section 7. Channelpropagations for both Hata and Rayleigh are mod-eled in section 8. Section 9 illustrates the simulationresults followed by conclusions.

Neural Network Handoff Algorithm in a Joint Terrestrial-HAPS Cellular System 165

Fig.1: The conceptual joint system model.

2. HYSTERESIS METHOD

A traditional handoff algorithm which is known asthe Hysteresis method uses relative signal strength asa main component in the handoff decision process. InFig. 2, the mobile is moving from the serving BS toanother BS. To have an ongoing call, we need to hand-off when the relative signal strength received from thetarget base rises above the Hysteresis margin dB. Itcorresponds to the distance at point C. For GSM sys-tem, handoff from one cell to another cell is decidedwhen [1]

RSS AV G = RSS AV GT − RSS AV GS ≥ h

where RSS AV G is the difference between theaveraged received signal strength from the tar-get cell (RSS AV GT ) and from the serving cell(RSS AV GS). However, the smaller h is, the morefrequent the unnecessary handoffs would happen. Inthis case, it might result in repeated handoffs betweenthe two BSs caused by rapid fluctuations in the re-ceived signal. It is so-called the ping-pong effect. Onthe other hand, having the larger hcan increase thedecision delay that might cause call dropping. Thereis therefore a tradeoff between the number of unnec-essary handoffs and decision delay.

Thus, we propose an improved handoff algorithmbased on RBF networks that consider mobile loca-tion, direction, averaged signal power and cell trafficintensity. Namely, the input to the neural network isnot only the received signal strengths, but it is alsothe direction of mobile moving to the target cell and

the traffic intensities of the neighbor cells. It makesthe proposed approach capable of reducing the hand-off rates and blocking rates as well as dropping ratescompared to the conventional Hysteresis method.

Fig.2: Location of a handoff decision in Hysteresis

3. RADIAL-BASIS FUNCTION NETWORK

In this section, we apply radial-basis functions(RBF) to the neural network as depicted in Fig. 3for making handoff decisions. The basic structureconsists of three layers; input layer, hidden layer andoutput layer [7]. The number of nodes used in the hid-den layer is 20. This number was found after trainingthe network and it made the errors converged. TheGaussian functions fully connect them to the five in-put nodes x = [x1, . . . , x5]. The output node is a lin-early weighted sum of the hidden unit outputs. Thereason that we have two outputs is to reduce numberof iterations of learning process. The outputs decidewhether the system needs a handoff or not. If y1 andy2 are 00, no handoff will be performed. If y1 and y2

are 11, then the system will handoff the mobile to thechosen base station.

The network is trained by the following pro-cedure to find the weight vector,wk1(n) =[wk(n), . . . , wk20(n)], starting at the iteration n = 0.

1. Initialize the center value µji(0) of the ith inputnode and the jth hidden node.

2. Initialize the span value σj(0) of the jth hiddennode.

3. Initialize the weight vector wk(0). Note that 1-3 are initialized using a uniform distribution between[0, 1] except w11(0) = w21(0) = 1.

4. Calculate the output of the hidden layer:zj = exp

( − ‖xi − µij(n)‖2/2σ2j (n)

)5. Calculate the output:

yk = R[ ∑M

j=i wkj(n)zj

], k = 1, 2; M = 20

where R[•] is a round operation.6. Calculate the error: ek = dk − yk where dk ∈

{0, 1} is the desired pattern.


Fig.3: Radial-basis function network

7. Update the weight: wkj(n + 1) = wkj(n) +ηwekzj where ηw is the learning rate of weight.

8. Update the center momentum:µij(n + 1) = µij(n) + ηµ

zj

σj(xi − µji(n))

∑ekwkj(n)

where ηµ is the learning rate of center.9. Update the span:

σj(n + 1) = σj(n) − 2ησzj

σj(n) lnzj

∑ekwkj(n)

where ησ is the learning rate of span.10. Repeat steps 4-9 until the mean square error

converges less than a small fixed number.The inputs to the neural network are listed below.1. x1 is the direction of a user. Under an assump-

tion that the mobile travels in unchanged direction, itcan be estimated using the MUSIC algorithm whichwill be described in the next section.

2. x2 is the signal strength of mobile received fromthe serving BS.

3. x3 is the signal strength of mobile received fromthe target BS. The received signal strength (RSS) isleveled as:low: −96 < RSS ≥ −87dBm and high: RSS > −87dBm.

4. x4 is the traffic intensity (TI) of the serving BS.5. x5 is the TI of the target BS. The TI is leveled

aslow: TI < 0.6591 Erlangs/Channelmedium: 0.6591 ≥ TI ≥ 0.76 Erlangs/ Channelhigh: TI > 0.76 Erlangs/ChannelNote that, we use the GOS (Grade of Service) in

the range 2-5% and the system has 20 channels todesign the TI levels. From the Erlang B table [10],TI equals to 13.1816 when the GOS is 2% and equalto 15.2 for 5%. Thus, the traffics per channel areequal to 0.6591 and 0.76, respectively.

Fig.4: Mean square errors in the trained network

The proposed handoff decision by the neural net-work depends on both RSSs and TIs of the serving(S) and target (T) BSs. We consider four cases asordered in Table 1 which are

1) Both the RSSs from the serving and target BSsare low.

2) The RSS from the serving BS is low while theRSS of the target BS is high.

3) Both the RSSs from the serving and target BSsare high.

4) The RSS from the serving BS is high while theRSS from the target BS is low.In each case, there are different handoff decisions(HO: handoff or NOHO: no-handoff) regarding to thelevels of traffic intensity.

Consider a particular case that the RSS from theserving cell is low and the RSS of target cell is highand their both intensities are high. In this severe sit-uation, the neural network decides definitely to hand-off. If the target cell chosen by using has no channelsavailable, we will consider among the adjacent cellsand HAPS. The cell that has the highest RSS will bethe cell to where mobile is handoffed.

In our implementation, 5000 samples are used fortraining the network and 500 samples are used to testthe network. For training, it is found that the errorsconverge to zero as shown in Fig. 4. Figure 5 illus-trates the good performance of the network for testingdata.

4. BACKPROPAGATION NEURAL NET-WORKS

The backpropagation (BP) structure shown in Fig.6 has the same number of nodes in each layer as Fig.3. Let xp= (xp1, xp2, xp3, . . . , xpN ) be a network in-put vector for N input nodes of pth the learning se-quence. In the nth iteration, vij(n) is the weight ele-ment connecting the ith input node and the jth hid-den node which is connected to the kth output node


Table 1: Handoff decision

by wjk(n). In the forward propagation, the input ofthe jth hidden node is

nethpj =N∑

i=1

vij(n)xpi

where vij(0) = 1. The output of the jth hidden nodesfor j = 1, . . . ,M can be obtained as

zpj = ϕ(nethpj)

where ϕ(x) = 11+e−x is the sigmoid function. Once

the output of the hidden node zpj is found, we can

Fig.5: Mean square errors in the tested network

compute the input of the output layer as

net0pk =M∑

j=1

wkj(n)zpj

where wkj(0) = 1. Note that the weight vectors areinitialized uniformly distributed between [0, 1]. Theoutput of the kth output node is a function as

ypk = R[ϕ(net0pk)]

In backward propagation, the procedure is startedfrom the output layer to the input layer to adjust theoutput and hidden weights given as follows:

δpk = ϕ′(net0pk)(dk − ypk)

where ϕ′(x) represents the differential. Next, up-

date the weight elements between the hidden and theoutput layers

∆wkj(n + 1) = ηδpkzpi + µ∆wkj(n).

where η is the learning rate and µ is the momentum.It is used to update the weight in the next iterationby adding

wkj(n + 1) = wkj(n) + ∆wkj(n + 1).

Likewise, the weights connecting the hidden and in-put layers are updated by using the following equa-tions.

δpj = ϕ′(nethpj)∑

k

δpkwkj(n + 1)

∆vji(n + 1) = ηδpjxpj + µ∆vji(n)

vji(n + 1) = vji(n) + ∆vji(n + 1).

After updating vji(n+1) and wkj(n+1), the forwardpropagation is repeated until the errors converge.

10000 samples are used for training and 500 sam-ples are used for testing. The convergence of erroris shown in Fig. 7. Figure 8 shows the mean squareerror resulting from testing the network.


Fig.6: Back-propagation neural network.

Fig.7: Mean square errors in the trained network

5. THE MUSIC ALGORITHM

Consider the received signals impinging on L uni-form linear array antennas which consist of the Msource signals corrupted by noise in the form of [9]:

X(t) =M∑

m=1

a(θm)sm(t) + η(t) = A(Θ)s(t) + η(t)

where A(Θ) = [a(θ1), . . . ,a(θM)] is an L×M steeringmatrix whose column is a steering vector a(θm) ofthe signal sm(t) having an unknown angle of arrivalθm. Θ = [θ1, . . . , θM] is an angle of arrival (AOA)vector of the M source signal vector denoted as s(t) =[s1(t), . . . , sM (t)]T with an assumption that M < L.The steering vector can be written as:

a(θm) = [1e−j2π dλ sinθm . . . e−j(L−1)2π d

λ sinθm ]T

where d/λ is the ratio of inter-element space of arrayto the signal wavelength. η(t) = [η1 (t), . . . , ηL(t)]T

Fig.8: Mean square errors in the tested network

is a noise vector which is temporally and spatiallywhite correlated.

The correlation matrix R of the array signal out-put can be written as

R = E[x(t)xH(t)].

An expression of R in terms of eigenvalues and theircorresponding eigenvectors is given as

R = UΛUH

where U = [u1, . . . ,uL]L×L is a unitary matrix andΛ is a diagonal matrix of real eigenvalues λi(λ1 ≥λ2 ≥ . . . ≥ λL > 0) corresponding to the eigenvectorui. Then, we decompose the correlation matrix Rinto a signal subspace: S = [u1, . . . ,uM] and a noisesubspace: N = [uM+1, . . . ,uL]. Thus, the matrix Rcan be rewritten as a sum of the two subspaces:

R = SΛsSH + NΛnNH.

As a result of orthogonality of the steering vectors tothe noise subspace, the MUSIC algorithm estimatesthe AOA vector of the signals by finding M peaks of

PMUSIC(θ) = |aH(θ)N|−2.

In practice, we cannot find the correlation matrix R,thus we use the sample-average estimate for N snap-shots as

R̂ =1N

N−1∑t=0

x(t)xH(t).

Figure 9 shows an AOA estimate of the ten sourcesignals corresponding to the ten peaks of MUSIC. TheFSK signal is used and defined as [13]

s(t) =

√2Es

Tcos[ωct + θ(t)]

where ωc is the carrier frequency, θ(t) the phase of thesignal, Es is the symbol energy and T is the symbol


period. In simulation, we assign T equal to 0.02 secs,ω1 and ω2 are equal to 100 and 200 Hz, Es = 1 dBand θ(t) is generated using the uniform distributionin the range of 0 − 2π. The number of antennas isused 30.

Fig.9: AOA estimates at 4◦, 63◦, 107◦, 137◦, 179◦,222◦, 256◦, 280◦, 328◦, 352◦.

6. TIMING ADVANCE

In the proposed handoff method, we need to knowthe mobile locations (d) to check if the mobiles are inan obstacle position or in the 3-cell junction. Posi-tioning a mobile station is divided into two categories:network-based and handset-based positioning. Weuse one of handset-based technique called the timingadvance (TA) [11-12] to determine where the mobileis.

TA is a measurement of the time required for thesignal to travel from the MS to the BS. The conceptof TA is used in order that the time-frames from eachMS arrive at the correct time slot when received bythe serving BS. By checking the position of the train-ing sequence transmitted from the MS on the uplink,the BS can calculate the TA value and send it back tothe MS in the downlink. Then, it enables the mobilesto know when to send the frame so that the signalsarrive at the BS in synchronism. For example, thevalue of TA in GSM system is an integer in a rangeof 0-63 according to 0 to 233 µs. It defines 0 is tobe no TA and 63 to be the maximum TA. Each TAnumber corresponds to different 550 m. radial dis-tance. It is reported to the MS every 480 ms duringthe connection.

In this paper, the TA is set in the range of 0 to 7according to 0 to 7 µs. These values are coded by 3bits. One bit of the TA represents a time difference of(7 µs)/7= 1 µs of the signal BS-MS-BS. The distanceper bit of TA is db = (3× 108 × 1µs)/2 = 150 m/TA.The MS-BS distance is a function of the integer TAas

d(TA) = TA × db

where TA = R[ν∆t/150] depending on the velocityof mobile ν and a time interval of measurement ∆t.

7. NEURAL NETWORK HANDOFF ALGO-RITHM

Figure 10 is a flow chart illustrating the proposedhandoff algorithm. The capacity of the system ismore efficient with the goodness of HAPS. First, wecheck if the mobile is staying in the obstacle positionor even at the corners of 3-cell junction by monitor-ing if the mobile distance is d < 800 or d ≥ 1000 andif the RSS is less than the threshold. The thresholdof the received signal strength (RSS drop) is at -82dBm before call is dropped. If the mobile has the sig-nal strength received from the serving BS less than-82 dBm, a handoff from the serving terrestrial BSto HAPS-BS is necessary to improve the call quality.If 800 ≤ d ≤ 1000, we measure the signal strengthof the mobile received from the current and adjacentcells every 0.5 sec. In order to decrease variation ofthe signals, an average of 10 data is calculated. Afterobtaining the averages of signal strength of currentand neighbor cells of x1, we can determine which cellsare most likely to be the target base stations (BST(i))by checking that the mobile receives the signal powerfrom which base stations radiate greater than -96dBm (RSS low TH). Included the knowledge of themobile directions obtained by the MUSIC algorithmand traffic intensities of the chosen BSs, the trainedneural network then makes a decision if the mobileneeds a handoff or not. In the case of deciding tohandoff, the algorithm will assign either case (1) towhich base station the mobile will be handoffed or(2) request a channel from HAPS. Note that HAPSgives higher priority to handoff-calls than new-callssuch that probabilities of blocking calls are low.

8. MOBILE RADIO PROPAGATION MODEL

8.1 Hata Model

The path loss in Hata model is expressed as [10]PL = 69.55+26.16logfc −13.82loghb −a(h)+(44.9−6.55loghm)logdwhere a(h) = 3.26(log11.75hm)2 − 4.97 for a carrierfrequency fc ≥ 300 MHz, hb is the height of antennaat base station (m) hm is the height of antenna atmobile (m), d is the distance between BS and MS(km). Accordingly, the strength of signal at a mobileis

RSS = P0 − PL

where P0 denotes the power radiated from the BS.

8.2 Shadow-Rayleigh Fading

In mobile radio channels, the Rayleigh distribution[4,10] is commonly used to model a flat fading chan-nel. This small-scale fading describes the statisticaltime varying nature of the received envelope. It is


Fig.10: Handoff algorithm using neural network.

well known that the sum of signals with the randomamplitudes and phases by no direct path between aMS and a BS obeys a Rayleigh distribution. The en-velope of the received signal strength rk at a MS-BSdistance d has a probability density function given by

p(rk) =rk

pkexp

( − r2k

2pk

), 0 ≤ rk ≤ ∞

where pk is the time-average power of the receivedsignal strength before envelope detection. Themean value of the Rayleigh distribution is E[rk] =√

πpk

2 . The received signal power with an arbitrarytransmitter-receiver separation distance is modeled asusing the path loss exponent from 2 to 6.

The shadow fading process is a Gaussian dis-tributed random process measured in decibels (dB).Under lognormal shadow fading, the autocorrelationis given as

RL(d) = σ2Lexp

(− |d|

d0

)

where σ2L and d0 are the variance and correlation

length of shadow fading, respectively. Accordingly,the power spectral density SL(v) can be found as

SL(v) =2d0σ

2L

1 + (2πvd0)2

where v is a spatial frequency. For a total traveleddistance D, the shadow fading process can be writtenas

L(d) =J∑

j=−J

√2

BDSL

( j

D

)cos

(2πjd

D+ φj

)

where

B =1

σ2LD

J∑j=−J

SL

( j

D

), J = Dvm.

Let vm be the maximum spatial frequency and φj bean identically independent uniform random process inrange of 0 − 2π. The discrete process Lk is obtainedby sampling the process at d = dk.

The received signal strength (RSS in dB) in thepresence of Rayleigh fading is in the form of

sk = 20log10rk

where rk is a Rayleigh distribution random variablewith parameter pk = (1/2)10mk/10. The power canbe calculated in dB as

mk = P0 − 10nlog10dk + Lk.

The mean and variance of are given as

s̄k = 10log10(2pk) − 10γ

ln10

= mk − 10γ

ln10

σ2sk

=50π2

3ln210

where γ ≈ 0.577216 is Euler’s Gamma.

9. SIMULATION RESULTS

The performance of the handoff algorithm usingthe RBF and BP neural networks is illustrated incomparison to the Hysteresis method. Simulationsare done for the system pictorially shown in Fig. 1.200 trials are run at a fixed traffic intensity and at afixed mean arrival time of the serving BS. The car-rier frequency is fc = 1800 MHz. The antenna heightis 60 m at BS and 1.5 m. at MS. The parameters ofcell environment and the statistical parameters of mo-bile are listed in Table 2-3. To evaluate the proposedmethod and Hysteresis rule, the handoff rate (H/Th),blocking rate (B/T) and dropping rate (D/Th) arecomputed at the serving BS (cell no. 1). H representsthe number of successful handoffed calls; Th repre-sents the total number of calls decided to handoff;B represents the number of blocked calls; Th repre-sents the total number of new calls; D represents thenumber of dropped calls. There are two cases in eachchannel model. First, we vary the traffic intensities inthe range of 0.65-0.77 Erlang/channel for acceptableGOS 2-5% of the cell No. 1 at the mean arrival timeequal to 1 min. In the other case, we vary the meanarrival time from 3-5 mins.

In Hata path loss propagation, Figs. 11-17 showthat handoff algorithm using the RBF network out-performs that using BP network and Hysteresis.Handoff rate, blocking rate and dropping rate result-ing from using RBF network are lowest in any case


of traffic intensities and mean arrival times as plottedin Figs. 11-16. Since we consider the signal strength,position, direction and traffic intensity, the unnec-essary handoffs is reduced as seen in Figs. 11 and14. The resulting blocking rates in Figs. 12 and15 have also low probabilities. As a result of thereduced unnecessary handoffs and HAPS, there aremore available channels for new calls to access the sys-tem. We also achieved the reduction of dropping ratesas shown in Figs. 13 and 16 because once the handoffrate decreases, chance of dropping is also less. Fig.17 depicts the number of handoffs occurring at thedifference between the signal power of mobile receiv-ing from the current BS and the minimum receivedsignal strength (-96 dBm) that the call will dropped.In fact, the less power difference, the more number ofhandoffs. Using the proposed algorithm, low numberof handoffs at low power difference is achieved. ForHysteresis rule, the power difference should higherthan 6 dB to avoid call-dropping.

In shadow-Rayleigh fading, we assign the Rayleighfading parameters in Table 3. Similarly, the resultsobtained from the proposed method using the RBFnetwork are the best among using BP network andusing Hysteresis as shown in Figs.18-24.

10. CONCLUSION

We propose an effective handoff algorithm basedon RBF network in high altitude platform station cel-lular systems. HAPS system can provide services tothe users staying at the corner of cells or at coveredarea influenced by shadowing. Besides the averagedreceived signal strength, user position and directionand traffic intensity are inputs of the neural networks.We use the RBF and BP to construct the networks.The MUSIC technique in the antenna array providesus an estimate of user direction. Timing advance usedin the GSM system is applied to this system for mo-bile positioning. Consequently, handoff rate, blockingrate and dropping rate decrease compared with thetraditional Hysteresis algorithm. The simulations areperformed in large scale channel propagation which isthe model of Hata path loss, and small scale channelpropagation in the process of Rayleigh fading. Be-sides the lower rates, we achieve low difference be-tween the mobile signal power from the target BS andthe minimum required power before dropping calls ata few needed numbers of handoffs.

References

[1] S. Ulukus and G. P. Pollini, “Handover Delay inCellular Wireless Systems,” IEEE InternationalConference on Communications, pp. 1370-1374,Jun. 1998.

[2] N. D. Tripathi, J. H. Reed and H. F. Van Landing-ham,“Pattern Classification Based Handoff UsingFuzzy Logic and Neural Nets,” IEEE Interna-

tional Conference on Communications, Vol. 3, pp.1733-1737, Jun. 1998.

[3] W. W. H. Yu and H. Changhua, “Resource Reser-vation in Wireless Networks based on PatternRecognition,” International Joint Conference onNeural Networks., Vol. 3, pp. 2264-2269, Jul.2001.

[4] R. Narasimhan and D. C. Cox, “A Handoff Algo-rithm for Wireless Systems Using Pattern Recog-nition,” 9th IEEE International Symposium onPersonal Indoor and Mobile Radio Communica-tions, Vol.

[5] K. D. Wong and D. C. Cox, “A Pattern Recogni-tion System for Handoff Algorithms,” IEEE Jour-nal on Selected Areas in Communications, Vol. 18,No. 7 pp. 1301-1312, Jul. 2000.

[6] K. D. Wong and D. C. Cox, “Two-State Pattern-Recognition Handoffs for Corner-Turning Situa-tions,” IEEE Transaction on Vehicular Technol-ogy, Vol. 50, No. 2, pp. 354-363, Mar. 2001.

[7] S. Haykin, Neural Networks: A ComprehensiveFoundation, Prentice Hall, 1999.

[8] S. Masumura and M. Nakagawa, “Joint Systemof Terrestrial and High Altitude Platform Station(HAPS) Cellular for W-CDMA Mobile Communi-cations,” IEICE Transaction on Communication,Vol. E85-B, No. 10, pp. 2051-2058, Oct. 2002.

[9] R. Suleesathira, “Close Direction of Arrival Es-timation for Multiple Narrowband Sources,” 7th

IEEE International Symposium on Signal Pro-cessing and Its Applications, Vol. 2, pp. 403-406,Jul. 2003.

[10] T. S. Rappaport, Wireless Communications:Principles and Practice, Prentice Hall, 2002.

[11] G. P. Yost and S. Panchapakesan, “Improvementin Estimation of Time of Arrival (TOA) fromTiming Advance (TA),” IEEE International Con-ference on Universal Personal Communications,Vol. 2, pp. 1367-1372, Oct. 1998.

[12] http://www.willassen.no/msl/node6.html[13] B. Sklar, Digital Communications Fundamentals

and Application, Prentice Hall, 2001.

Raungrong Suleesathira received theB.S. degree in 1994 from Kasetsart Uni-versity, Bangkok, Thailand and the M.S.and Ph.D. degrees from University ofPittsburgh, PA, in 1996 and 2001, re-spectively, all in Electrical Engineering.Since 2001, she has been on the fac-ulty of the Department of Electronicsand Telecommunication Engineering atKing Mongkut’s University of Technol-ogy Thonburi, Thailand where she is

now an Assistant Professor. Her current research interests arein the areas of smart antennas and wireless communications.


Sunisa Kunarak received the B.S.degree in Electronics and Telecommu-nication Engineering and M.S. degreein Electrical Engineering from KingMongkut’s University of TechnologyThonburi, Thailand in 2002 and 2004,respectively. Currently, she works asa communication engineer for process-ing mobile signals in subways at BombarDier Transportation (Signal) Thailand.

Table 2: Cell Structure Parameters

Parameters value

cell radius 1000 m

number of channels per cell 20 channels

number of channels in HAPS 20 channels

low threshold of received signal -96 dB

strength (RSS low TH)

received signal strength for call -82 dB

dropping (RSS drop)

Power Transmitted p0 40 dBm

Hysteresis level (h) 6 dB

Table 3: Mobile Parameters

variable distribution interval

initial position in the Uniform [-10,10]

horizontal (m)

initial position in the Uniform [-8,8]

vertical (m)

number of users in Uniform [100,400]

each cell

mean arrival time of Poisson 1 min

new call

direction of mobile Uniform [0,2π ]

average time per call Exponential 120 sec

average mobile speed Normal N(60,10)

km/hr

Table 4: Rayleigh Fading Parameters

Parameters value

Minimum Base Station Separation 1600 m

Exponent of Distance Dependence (n) 4

Standard Deviation of Shadow Fading σL 8 dB

Correlation Length of Shadow Fading d0 20 m

Number of terms for Shadow Fading Process 321

Maximum Spatial Frequency 0.1

Fig.11: Handoff rate versus traffic intensity in Hata

Fig.12: Blocking rate versus traffic intensity in Hata

Fig.13: Dropping rate versus traffic intensity inHata


Fig.14: Handoff rate versus mean arrival time inHata

Fig.15: Blocking rate versus mean arrival time inHata

Fig.16: Dropping rate versus mean arrival time inHata

Fig.17: Signal difference versus number of handoffsin Hata

Fig.18: Handoff rate versus traffic intensity inRayleigh

Fig.19: Blocking rate versus traffic intensity inRayleigh


Fig.20: Dropping rate versus traffic intensity inRayleigh

Fig.21: Handoff rate versus mean arrival time inRayleigh

Fig.22: Blocking rate versus mean arrival time inRayleigh

Fig.23: Dropping rate versus mean arrival time inRayleigh

Fig.24: Signal difference versus number of handoffsin Rayleigh

Neural Network Handoff Algorithm in a Joint Terrestrial-HAPS Cellular System

Documents

Transcript of Neural Network Handoff Algorithm in a Joint Terrestrial-HAPS Cellular System